reintroduce invsqrt and modernize it.

[Ka-zam]

This PR gets rid of the Q_rsqrt implementation and the puppet. Modern CPUs don't use that and it's hacky at best. Fails on negative inputs.

Tolerance is tightened from 1e-02 to 1e-6.

Last kernel before release!

RUN_VOLK_TESTS: volk_32f_invsqrt_32f(vlen=131071, iter=3000, tol=1e-06)
arch                       |         time |     throughput |    max_rel |
---------------------------+--------------+----------------+------------+
generic                    | 1797.7696 ms |    1749.8 MB/s |          - |
a_avx                      |   22.2048 ms |  141673.2 MB/s |    2.4e-07 |
a_avx512f                  |   15.3338 ms |  205156.5 MB/s |    2.1e-07 | *
a_sse                      |   38.4642 ms |   81785.7 MB/s |    2.4e-07 |
u_sse                      |   37.6666 ms |   83517.7 MB/s |    2.4e-07 |
u_avx                      |   19.6460 ms |  160125.7 MB/s |    2.4e-07 |
u_avx512f                  |   15.7183 ms |  200137.9 MB/s |    2.1e-07 | *
Best aligned arch          | a_avx512f (117.24x)
Best unaligned arch        | u_avx512f (114.37x)
--------------------------------------------------------------------------------

RUN_VOLK_TESTS: volk_32f_invsqrt_32f(vlen=131071, iter=597, tol=1e-06)
arch                       |         time |     throughput |    max_rel |
---------------------------+--------------+----------------+------------+
generic                    |  689.8667 ms |     907.4 MB/s |          - |
neon                       |   44.6217 ms |   14029.5 MB/s |    2.0e-07 |
neonv8                     |   43.7900 ms |   14295.9 MB/s |    2.0e-07 | *
Best aligned arch          | neonv8 (15.75x)
Best unaligned arch        | neonv8 (15.75x)
--------------------------------------------------------------------------------

[jdemel]

Mostly LGTM.

Just two questions for clarification =D

[jdemel]

Thanks for the tests and clarifications =D LGTM!